The smallest member of the family: an 80 MB, int8-quantized static embedding model for educational metadata, built for the tightest deployments (edge, embedded, memory-limited). It is the int8 variant of m2v-gte-256-edu — same Model2Vec distillation of Alibaba-NLP/gte-multilingual-base (mean pooling, 256-d, ~70k German education keyword vocabulary), then quantized to int8.

numpy-only inference, no PyTorch. The base is multilingual; the domain vocabulary is primarily German.

Quantizing the 256-d table from float32 to int8 cuts the model to a quarter of the size, at a small, measured cost on our German subject-classification benchmark (25k items, 47 labels, honest held-out test):

So int8 costs about 1 macro-F1 point for a 4× smaller model. (Note: the per-vector cosine similarity between int8 and float32 embeddings is ~0.9997 — very high — yet the downstream classifier is slightly more sensitive than that suggests, hence the ~0.01 drop. Measure, don't assume.) For memory/bandwidth-constrained applications this is usually a good deal; where accuracy matters more, use the float32 sibling.

Fast embedding of (mostly German) educational metadata — titles, descriptions, keywords — in the smallest possible footprint: edge devices, embedded systems, serverless/cold-start, or anywhere an 80 MB model beats a 321 MB one. For classification, clustering, semantic search and retrieval.

Or via Sentence Transformers:

Model2Vec passes a vocabulary through a sentence-transformer, reduces dimensionality with PCA, and weights the token embeddings with SIF weighting. At inference it takes the (weighted) mean of the static token vectors of a text — no transformer forward pass.

Inherits the base model's license (gte-multilingual-base: Apache-2.0).

@article{minishlab2024model2vec,
  author = {Tulkens, Stephan and {van Dongen}, Thomas},
  title = {Model2Vec: Fast State-of-the-Art Static Embeddings},
  year = {2024},
  url = {https://github.com/MinishLab/model2vec}
}